Skip to content

[R-meta] possible miscalculation of Cook’s distances

8 messages · Reza Norouzian, Wolfgang Viechtbauer, Antonina Dolgorukova

#
Dear Dr. Viechtbauer and all,

I have a multilevel data structure (experiments nested within studies) are
use rma.mv to calculate an overall effect estimate. The next step requires
sensitivity analysis on the experiment-level data. For detecting outliers
I've used standardized (deleted) residuals and for detecting influential
experiments I've used Cook?s distance. However, the last test provides
contradictory results.

The forest plot indicates that the 1st experiment may be an outlier, and
standardized (deleted) residuals confirm this. But according to Cook?s
distance plot, the 2nd experiment is influential. It seems that there may
be a miscalculation of Cook?s distance since I can easily reproduce this
issue (in one study one of the experiments have to provide a much larger ES
than the other) also if use the model with random = ~ 1 | experiment, the
1st experiment is influential, not the second.

Could you please clarify is this a bug or a feature of the cooks.distance()
function? Maybe it does not work properly with rma.mv objects?

  ## Reproducible Example

# data frame
  dat <- data.frame(study=c(1,1,2,3,3,3), experiment=c(1:6),
                    yi=c( 68, 18, 31,20,10,26),
                    vi=c(60,32, 15, 19, 41, 82))

# multilevel model
res.ml <- rma.mv(yi, vi,
                 random = ~ 1 | study/experiment,
                 data=dat,
                 slab = paste("Study ", study,", ", "Experiment ",
experiment, sep = ""))

# forest plot examination  indicates that the 1st experiment may be an
outlier
forest(res.ml,
       header = "Study and Experiment ID")

# standardized (deleted) residuals confirm this
rst <- rstudent(res.ml)
plot(NA, NA, xlim=c(1, res.ml$k), ylim=c(-3,5),
     xlab="Study and Experiment ID", ylab="Standardized (Deleted) Residual",
     xaxt="n", las=1)
axis(side=1, at=1:res.ml$k, labels=rst$slab)
abline(h=c(-1.96,1.96), lty="dotted")
abline(h=0)
points(1:res.ml$k, rst$z, type="o", pch=19)

# according to Cook?s distance plot, the 2nd experiment is influential
cooksd <- cooks.distance(res.ml)
plot(1:length(cooksd), cooksd, ylim=c(0,2), type="o", pch=19, las=1,
xaxt="n",
     xlab="Experiment ID", ylab="Cook's Distance")
axis(1, 1:length(cooksd), labels=names(cooksd))
abline(h=4/length(cooksd), lty="dotted", lwd=2)

sessionInfo()
R version 4.1.2 (2021-11-01)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 10 x64 (build 19042)
[1] metafor_3.1-43 metadat_1.0-0  Matrix_1.3-4
#
Dear Antonina,

An effect estimate whose standardized deleted residual falls beyond
+/-1.96 doesn't necessarily need to have a cook's distance
(influence), and/or a hat value (if additional moderators are used)
that are likewise extreme.

As a general proposition, it may be methodologically more reasonable
to simultaneously consult all these indices to flag an extreme effect
estimate.

That said, in your case, it does seem that the estimate from study 1,
experiment 1 has a large standardized deleted residual as well as a
large Cook's distance relative to that of other effect estimates. So,
I don't think there is any bug in any of the metafor functions you
used.

Below is a bit of code to catch that outlying effect estimate.

Kind regards,
Reza

dat <- data.frame(study=c(1,1,2,3,3,3), experiment=c(1:6),
                  yi=c( 68, 18, 31,20,10,26),
                  vi=c(60,32, 15, 19, 41, 82))

res.ml <- rma.mv(yi, vi,
                 random = ~ 1 | study/experiment,
                 data=dat,
                 slab = paste("Study ", study,", ", "Experiment ",
                              experiment, sep = ""))

dat <- transform(dat, obs = res.ml$slab)

outlier_limit <- qnorm(.975)

cook <- cooks.distance.rma.mv(res.ml,
                              reestimate = FALSE)

st_del_res_z <- rstudent.rma.mv(res.ml,
                                reestimate = FALSE)$z

cook_limit <- max(mean(range(cook)), boxplot.stats(cook, coef = 1.5)$stats[5])

i <- abs(st_del_res_z) > outlier_limit
k <- cook > cook_limit
oo <- which(i & k)

LL <- names(oo)

removed <- subset(dat, obs %in% LL)
new_dat <- subset(dat, !obs %in% LL)







On Thu, Feb 3, 2022 at 3:36 PM Antonina Dolgorukova
<an.dolgorukova at gmail.com> wrote:
#
Dear Reza,

Thank you very much for your reply. I do understand that not all outliers
have to be influential and that not all influential cases have to be
outliers. But thank you for mentioning this and for the code provided. I
did notice you set the reestimate = FALSE. This raises two questions,
actually. It would be great if you help with them.

1) Is there any explanation why cooks.distance.rma.mv(res.ml) and
cooks.distance.rma.mv(res.ml , reestimate = FALSE) give completely
different results about the experiments 1 and 2 (and very similar for the
remaining experiments)?

According to cooks.distance.rma.mv(res.ml), the Study 1, Experiment 2 is
influential, but
according to  cooks.distance.rma.mv(res.ml , reestimate = FALSE), Study 1,
Experiment 1 is influential


#the code illustrating my question

dat <- data.frame(study=c(1,1,2,3,3,3), experiment=c(1:6),
                  yi=c( 68, 18, 31,20,10,26),
                  vi=c(60,32, 15, 19, 41, 82))

res.ml <- rma.mv(yi, vi,
                 random = ~ 1 | study/experiment,
                 data=dat,
                 slab = paste("Study ", study,", ", "Experiment ",
experiment, sep = ""))

cook_with_reest <- cooks.distance.rma.mv(res.ml)
cook_without_reest <- cooks.distance.rma.mv(res.ml, reestimate = FALSE)

plot(1:length(cook_with_reest), cook_with_reest, ylim=c(0,2), type="o",
pch=19, las=1, xaxt="n",
     xlab="Study and Experiment ID", ylab="Cook's Distance")
points(cook_without_reest, type="o", pch=19, col = "red")
axis(1, 1:length(cook_with_reest), labels=names(cook_with_reest))
abline(h=4/length(cook_with_reest), lty="dotted", lwd=2)
legend("top", pch=19, col=c("black","red"), lty="solid",
       legend=c("reestimate = TRUE","reestimate = FALSE"), bty="n")



2) And could you please explain what is the rationale to set reestimate =
FALSE? According to the metafor documentation:
"Doing so only yields an approximation to the Cook?s distances that ignores
the influence of the ith case on the variance/correlation components"


Sincerely,
Antonina
On Fri, Feb 4, 2022 at 3:29 AM Reza Norouzian <rnorouzian at gmail.com> wrote:

            

  
    
#
Dear Antonina,

There was no rationale, I just wanted to indicate that there is no bug
in the function (and save myself a tiny bit of time). The default
reestimate = TRUE is certainly to be preferred given that your model
does include random-effects and, as the documentation correctly
mentions, the influence of each effect estimate on the estimates of
heterogeneity and correlation can only be examined, if you set
reestimate = TRUE.

For instance, if you remove the random-effects from your initial model
(res.ml), then, you'll see that the use of reestimate = TRUE or FALSE
has no effect. In both cases, only the effect estimate associated with
study 1, experiment 1 is influential. Thus, this shows that the use of
reestimate = FALSE in your initial model essentially ignored the
influence of each effect estimate on the estimates of heterogeneity.

In models with complex random-effects structure esp. fit to large
datasets, setting reestimate = TRUE would take a good chunk of time.
In such cases, some (myself included) may be tempted to set reestimate
= FALSE hoping that the approximation will be close enough.

Kind regards,
Reza

res.ml2 <- update.rma(res.ml, random = NULL)

cook_with_reest2 <- cooks.distance.rma.mv(res.ml2)
cook_without_reest2 <- cooks.distance.rma.mv(res.ml2, reestimate = FALSE)

plot(cook_with_reest2,type="b")
plot(cook_without_reest2,type="b")

On Fri, Feb 4, 2022 at 12:19 AM Antonina Dolgorukova
<an.dolgorukova at gmail.com> wrote:
#
Just as a brief follow-up with respect to the calculations: They are correct. Continuing with your code, Antonina, we can easily manually compute the Cook's distances here:

cooksd[1:2]

sub <- rma.mv(yi, vi,
              random = ~ 1 | study/experiment,
              data=dat, subset=-1)

(coef(res.ml) - coef(sub))^2 / vcov(res.ml)

sub <- rma.mv(yi, vi,
              random = ~ 1 | study/experiment,
              data=dat, subset=-2)

(coef(res.ml) - coef(sub))^2 / vcov(res.ml)

As you will see, these match up.

With these data, removal of a single row can have a large impact on the variance components, which can lead to considerable changes in the pooled estimate.

Best,
Wolfgang
#
Dear Wolfgang,

Thank you very much for confirming the calculations!! Tha's great that they
are correct! Actually, by "miscalculation", I have rather meant the order,
not the numbers.

The thing that confuses me a lot is that according to forest plot
examination and standardized (deleted) residuals, the 1st experiment is an
outlier. But according to Cook?s distance plot, the 2nd experiment is
influential. Looking at the forest plot, it is really hard to believe that
the removal of the 2nd experiment can affect the pooled estimate (its mean
is close to the overall effect estimate and CI does overlap (almost
included) with CI of the overall effect).
[image: image.png]
[image: image.png]


Sincerely,
Antonina

On Fri, Feb 4, 2022 at 7:17 PM Viechtbauer, Wolfgang (SP) <
wolfgang.viechtbauer at maastrichtuniversity.nl> wrote:

            

  
    
#
This is due to the way the estimates are weighted, which is a consequence of the variance components. In the full data, all estimates receive roughly equal weights:

weights(res.ml, type="rowsum")

On the other hand, when you remove experiments 1 or 2, the VCs change such that the weights also change noticeably:

weights(sub, type="rowsum")

Now, experiment 1 or 2 (whichever is still in the subset) receives more weight and so does experiment 3. So, removal of experiment 2 leads to more weight for experiment 1 which also has an extreme estimate, which affects the pooled estimate considerably. Hence, the large Cook's distance.

Best,
Wolfgang
1 day later
#
Dear Wolfgang,

Thank you for the clarification. It really helps. I didn?t know that with
Cook's distance one may not only identify cases that are influential itself
but also cases, whose removal would affect the pooled estimate indirectly
(through the reallocation of weights of the remaining cases). If this were
the only criterion in my analysis, after removal of that one case I might
get pooled estimate biased towards an outlier. It?s good we have more
options for data examination and do use them simultaneously.

Your help is much appreciated!
Best regards,
Antonina

On Sat, Feb 5, 2022 at 5:59 PM Viechtbauer, Wolfgang (SP) <
wolfgang.viechtbauer at maastrichtuniversity.nl> wrote: